20 October 2026
A new chart tracks two and a half years of uneven AI progress
First reported
One Useful Thing ran this on .
- A METR benchmark measuring the length of tasks AI can complete has doubled eight times since GPT-4, a model from OpenAI.
- Answers to standard benchmark tests have become much cheaper, and about 180 models have now surpassed GPT-4.
- Gains are uneven: models are strong at some tasks and weak at others, rather than improving evenly across the board.
How it was covered
One Useful ThingEthan Mollick
Long-horizon task benchmark trends, falling AI costs, model leaderboard changes, the jagged frontier, agentic work, and his new book